Have been running the BOINC client under
Mandrake Linux 10.0, 32bit on an
Athlon64 3200+
512 MB DDR400
since about 1900 UTC/GMT Saturday. Through three trickles, the averge sec/ts has been 2.11 (~6:22 per trickle).

This is quite a bit slower than the classic Windows client which averaged almost exactly 2 sec/ts (trickle times from 5:55 to 6:08 for 5 runs). This is unfortunately works out to a full day longer for 3 phase model completion. Same hardware now as then.

I know the Intel CPUs appear to be running faster with the BOINC client (Intel compiler?) so I'm not sure a comparison is possible there. But has anyone with an Athlon who has switched from Windows classic CPDN to BOINC Linux seen a similar slowdown? Thanks.

although I don't have enough data yet I wouldn't be surprised if the Intel compilers are now favoring Pentium's? It certainly seems so from the <A HREF="http://climateapps2.oucs.ox.ac.uk/cpdnboinc/cpu.html">crude CPU performance page</a> (where a Pentium Linux box is in the lead).

From "A/B" tests here on the same machine (alternately booting into Linux &amp; WinXP) it seems that Linux is actually 10%-15% faster on the same exact machine.

I just booted into my Mandrake AMD64-bit Linux to try out -- I am getting 2.38 seconds per timestep here versus 2.42 s/TS in WinXP, so it's not as dramatic an increase on my home PC than the ones at work.

&gt; although I don't have enough data yet I wouldn't be surprised if the Intel
&gt; compilers are now favoring Pentium's? It certainly seems so from the <A> HREF="http://climateapps2.oucs.ox.ac.uk/cpdnboinc/cpu.html"&gt;crude CPU
&gt; performance page</a> (where a Pentium Linux box is in the lead).
&gt;
&gt; From "A/B" tests here on the same machine (alternately booting into Linux
&gt; &amp; WinXP) it seems that Linux is actually 10%-15% faster on the same exact
&gt; machine.
&gt;
&gt; I just booted into my Mandrake AMD64-bit Linux to try out -- I am getting 2.38
&gt; seconds per timestep here versus 2.42 s/TS in WinXP, so it's not as dramatic
&gt; an increase on my home PC than the ones at work.
&gt;
I guess I'll start a client in Windows and see if it is really slower there as well. It's possible what you are seeing, maybe, is the run to run speed variability due to different parameters sets. Unless you've seen this across several models run in parallel? Or, maybe it runs relatively better under AMD64-bit rather than 32bit?

&gt; From "A/B" tests here on the same machine (alternately booting into Linux
&gt; &amp; WinXP) it seems that Linux is actually 10%-15% faster on the same exact
&gt; machine.
&gt;
&gt; I just booted into my Mandrake AMD64-bit Linux to try out -- I am getting 2.38
&gt; seconds per timestep here versus 2.42 s/TS in WinXP, so it's not as dramatic
&gt; an increase on my home PC than the ones at work.
&gt;
&gt;
Curiousity got the better of me so I installed another BOINC client in WinXP on the same machine. It's getting 2.08 in WinXP vs. 2.11 in Linux. But that may just be model parameter differences. Still, even in WinXP, it's about 8 minutes slower per trickle than the slowest classic model run on this PC.

What flags/compiler options did you compile the client with?